Paper: arXiv 2412.18405

Authors: Jakub Michańków, Paweł Sakowski, Robert Ślepaczuk

Abstract

Regardless of the selected asset class and the level of model complexity (Transformer versus LSTM versus Perceptron/RNN), the GMADL loss function produces superior results than standard MSE-type loss functions and has better numerical properties in the context of optimization than MADL. Better results mean the possibility of achieving a higher risk-weighted return based on buy and sell signals built on forecasts generated by a given theoretical model estimated using the GMADL versus MSE or MADL function. In practice, GMADL solves the problem of selecting the most preferable feature in both classification and regression problems, improving the performance of each estimation. What is important is that, through additional parameterization, GMADL also solves the problem of optimizing investment systems on high-frequency data in such a way that they focus on strategy variants that contain fewer transactions so that transaction costs do not reduce the effectiveness of a given strategy to zero. Moreover, the implementation leverages state-of-the-art machine learning tools, including frameworks for hyperparameter tuning, architecture testing, and walk-forward optimization, ensuring robust and scalable solutions for real-world algorithmic trading.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced mathematical derivations for a novel generalized loss function (GMADL) with multiple parameters and exponential smoothing, indicating high complexity. It also demonstrates strong empirical rigor through comprehensive backtesting on high-frequency data (Bitcoin, Crude Oil), hyperparameter tuning, walk-forward optimization, and risk-weighted return comparisons.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve ML models for HFT<br>by reducing overfitting & transaction costs"] --> B["Data Input: High-Frequency Market Data"]
  B --> C{"Methodology: Compare Loss Functions"}
  C --> D["Model 1: Standard MSE Loss"]
  C --> E["Model 2: MADL Loss"]
  C --> F["Model 3: GMADL Loss"]
  
  subgraph Computational Process ["Model Estimation & Optimization"]
      direction LR
      G["Hyperparameter Tuning"] --> H["Walk-Forward Optimization"]
      H --> I["Signal Generation"]
  end
  
  D --> G
  E --> G
  F --> G
  
  G --> J["Key Findings/Outcomes"]
  
  J --> K["GMADL yields superior<br>risk-weighted returns"]
  J --> L["Optimizes for fewer transactions<br>reducing costs significantly"]
  J --> M["Solves feature selection<br>in both regression & classification"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes the Generalized Mean Absolute Directional Loss (GMADL) function as a new loss for training machine learning models in high‑frequency algorithmic trading.

Source sentence
Regardless of the selected asset class and the level of model complexity (Transformer versus LSTM versus Perceptron/RNN), the GMADL loss function produces superior results than standard MSE-type loss functions and has better numerical properties in the context of optimization than MADL.

Assumes 3

  • High‑frequency data entail substantial transaction costs that must be accounted for in the loss function. market
    Source sentence
    What is important is that, through additional parameterization, GMADL also solves the problem of optimizing investment systems on high-frequency data in such a way that they focus on strategy variants that contain fewer transactions so that transaction costs do not reduce the effectiveness of a given strategy to zero. GMADL function - adaptability and scalability Taking into account the existence of transaction
  • The loss function should be differentiable enough to allow gradient‑based optimization. model
    Source sentence
    One of the primary challenges lies in gradient optimization during the learning process. The reliance on absolute and sign functions introduces non-differentiability at zero, creating complications for smooth gradient-based training.
  • Directional accuracy and magnitude of returns are more relevant for investment performance than pure point‑forecast error. statistical
    Source sentence
    We highlighted that standard error metrics fail to capture the effectiveness of forecasts for generating investment signals. These metrics focus only on forecast accuracy, penalizing all errors equally, regardless of direction, or consider only directional accuracy, ignoring error magnitude.

Data

  • Bitcoin daily simple returns crypto-onchain unknown
    Source sentence
    Then, in the Results section (in the second version of this paper), we show an empirical comparison of the use of the MADL function and the classical forecast error function (MAE) in two time series: Bitcoin and Crude Oil daily simple returns.
  • Crude Oil daily simple returns other unknown
    Source sentence
    Then, in the Results section (in the second version of this paper), we show an empirical comparison of the use of the MADL function and the classical forecast error function (MAE) in two time series: Bitcoin and Crude Oil daily simple returns.

Costs

Not applicable The paper reports no trading or execution result; the cost the extractor found is a model parameter.

The paper-text extraction found a cost parameter inside a model (The paper discusses parameterizing the loss (a and b) to reduce the number of transactions and mentions transaction costs as a factor, but does not embed a specific cost model in the headline results.); the classifier found no reported trading result, so costs are not applicable.

Finds 5

  • GMADL yields superior results compared with standard MSE‑type loss functions and the original MADL across asset classes and model complexities. performance
    Source sentence
    Regardless of the selected asset class and the level of model complexity (Transformer versus LSTM versus Perceptron/RNN), the GMADL loss function produces superior results than standard MSE-type loss functions and has better numerical properties in the context of optimization than MADL.
  • Empirical tests on Bitcoin and Crude Oil daily returns show that GMADL outperforms the classical MAE loss. performance
    Source sentence
    Then, in the Results section (in the second version of this paper), we show an empirical comparison of the use of the MADL function and the classical forecast error function (MAE) in two time series: Bitcoin and Crude Oil daily simple returns.
  • The additional parameterization (a, b) enables the loss to focus on strategies with fewer transactions, mitigating the impact of transaction costs. theoretical
    Source sentence
    What is important is that, through additional parameterization, GMADL also solves the problem of optimizing investment systems on high-frequency data in such a way that they focus on strategy variants that contain fewer transactions so that transaction costs do not reduce the effectiveness of a given strategy to zero. GMADL function - adaptability and scalability Taking into account the existence of transaction
  • Using exponential smoothing with parameters a and b makes GMADL more differentiable and stabilizes optimization in diverse market environments. theoretical
    Source sentence
    Using exponential smoothing along with parameters a and b, we have shown how GMADL offers a differentiable and adaptable framework, thus enhancing the stability and efficiency of optimization processes in diverse market environments.
  • GMADL can be scaled to high‑frequency, minute‑by‑minute or tick‑level forecasting contexts while maintaining adaptability across asset classes. performance
    Source sentence
    The result is a highly scalable loss function that not only caters to traditional daily or weekly trading horizons but also excels in granular, minute-by-minute, or even tick-level forecasting contexts. 3.5.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 37 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (103)

  1. The lorentz transformation and absolute time (1953) — link
  2. The theory of a general quantum system interacting with a linear dissipative system (1963) — link
  3. New evidence on the implied-realized volatility relation (2002) — link
  4. Forecasting Volatility in the Financial Markets (2007) — link
  5. The relationship between implied and realized volatility of S&P 500 index (2003) — link
  6. Implied Volatilities as Forecasts of Future Volatility, Time-Varying Risk Premia, and Returns Variability (2001)
  7. The relation between implied and realized volatility (1998) — link
  8. Volatility forecasting in the framework of the option expiry cycle (1999) — link
and 95 more
  1. The Relationship between Implied and Realized Volatility in the Danish Option and Equity Markets (1999) — link
  2. Information Content of Option Implied Volatility : Evidence from the Indian Market (2008) — link
  3. The relationship between implied and realized volatility: evidence from the Australian stock index option market (2009) — link
  4. Predicting Volatility in the Foreign Exchange Market (1995) — link
  5. Econometric Computing with HC and HAC Covariance Matrix Estimators (2004) — link
  6. Implementing a class of structural change tests: An econometric computing approach (2006) — link
  7. Object-oriented Computation of Sandwich Estimators (2006) — link
  8. The Informational Content of Implied Volatility (1993) — link
  9. Machine Learning Methods in Algorithmic Trading Strategy Optimization–Design and Time Efficiency (2019)
  10. Is smart beta really smart? (2014)
  11. How Smart are Smart Beta ETFs? Analysis of Relative Performance and Factor Exposure (2015)
  12. Do smart beta ETFs deliver persistent performance? (2020)
  13. Portfolio theory in terms of partial covariance (2016)
  14. Patterns in financial markets: Dynamic time warping (2012)
  15. An empirical study on the differences between synthetic and physical ETFs (2015)
  16. Tracking ability of ETFs: physical versus synthetic replication (2014)
  17. Tracking S&P 500 index funds (2001)
  18. Robo-Advisors: Machine Learning in Trend-Following ETF Investments (2020)
  19. Similarity learning for time series classification (2016) — link
  20. Guide to pragmatic replication (2019)
  21. A novel dynamic asset allocation system using Feature Saliency Hidden Markov models for smart beta investing (2021)
  22. Redefining Indexing Using Smart Beta Strategies (2017)
  23. The asset manager’s dilemma: How smart beta is disrupting the investment management industry (2016)
  24. Fat Tailed Factors (2020) — link
  25. Competition for Attention in the ETF Space (2021)
  26. Alternative Beta: Redefining Alpha and Beta
  27. Making Your Smart Beta Even Smarter (2014)
  28. Tail Risk Mitigation with Managed Volatility Strategies (2017)
  29. Ten misconceptions about smart beta (2016)
  30. TailCoR (2020)
  31. Managing Beta Exposures in Portable Alpha Strategies
  32. Tail Risk of Smart Beta Portfolios: An Extreme Value Theory Approach (2014)
  33. Predicting prices of S&P500 index using classical methods and recurrent neural networks (2020)
  34. Application of machine learning in quantitative investment strategies on global stock markets (2021)
  35. Predicting stock prices using LSTM (2017)
  36. A LSTM-based method for stock returns prediction: A case study of China stock market (2015)
  37. Improving trading technical analysis with tensorflow long short-term memory (lstm) neural network (2019)
  38. Nvidia’s Stock Returns Prediction Using Machine Learning Techniques for Time Series Forecasting Problem (2020)
  39. Improving stock closing price prediction using recurrent neural network and technical indicators (2018)
  40. Forex exchange rate forecasting using deep recurrent neural networks (2020)
  41. Deep learning with long short-term memory networks for financial market predictions (2018)
  42. Using neural networks to enhance technical trading rule returns: A case with KLCI (2016)
  43. Deep leaming for stock market prediction using technical indicators and financial news articles (2018)
  44. Forecasting price movements using technical indicators: Investigating the impact of varying input window length (2017)
  45. All that glitters is not gold: Comparing backtest and out-of-sample performance on a large cohort of trading algorithms (2016)
  46. What to Look for in a Backtest (2013)
  47. The probability of backtest overfitting (2016)
  48. Backtest overfitting in financial markets (2016)
  49. Portfolio of Global Futures Algorithmic Trading Strategies for Best Out-of-Sample Performance (2016)
  50. Algorithmic trading: winning strategies and their rationale (2013)
  51. Quantitative trading: how to build your own algorithmic trading business (2021)
  52. Machine Learning for Algorithmic Trading: Predictive models to extract signals from market and alternative data for systematic trading strategies with Python (2020)
  53. A high-frequency algorithmic trading strategy for cryptocurrency (2020)
  54. Redes Neurais | LSTM (2019) — link
  55. Applying Hybrid ARIMA-SGARCH in Algorithmic Investment Strategies on S&P500 Index (2022)
  56. Applying Hurst Exponent in pair trading strategies on Nasdaq 100 index (2021) — link
  57. Deep Learning with Python, 2nd ed. (2021)
  58. Long Short - Term Memory (1997) — link
  59. Learning to Forget : Continual Prediction with LSTM (2000) — link
  60. Learning to forget: continual prediction with LSTM (1999) — link
  61. Multi Factor Stock Selection Model Based on LSTM (2018) — link
  62. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling (2014) — link
  63. Deep learning with long short-term memory networks for financial market predictions (2018) — link
  64. Deep Learning (2016)
  65. Investment Strategies that Beat the Market . What Can We Squeeze from the Market ? (2018) — link
  66. AT - LSTM : An Attention -based LSTM Model for Financial Time Series Prediction (2019) — link
  67. Momentum and contrarian effects on the cryptocurrency market (2019) — link
  68. Deep Learning (2016)
  69. Adam: A Method for Stochastic Optimization (2017) — link
  70. Hyperband: A Novel Bandit - Based Approach to Hyperparameter Optimization (2018) — link
  71. KerasTuner (2019)
  72. Artificial neural networks architectures for stock price prediction: Comparisons and applications (2016)
  73. Multi Factor Stock Selection Model Based on LSTM (2018) — link
  74. Deep Learning for Stock Selection Based on High Frequency Price - Volume Data (2019) — link
  75. Deep Learning in Finance (2018) — link
  76. Modified support vector machines in ÿnancial time series forecasting (2002)
  77. Dynamic financial distress prediction with concept drift based on time weighting combined with Adaboost support vector machine ensemble (2017) — link
  78. Financial time series prediction using least squares support vector machines within the evidence framework (2001) — link
  79. Using deep learning to detect price change indications in financial markets (2017) — link
  80. A New Kernel of Support Vector Regression for Forecasting High - Frequency Stock Returns (2016) — link
  81. Machine-learning classification techniques for the analysis and prediction of high-frequency stock direction - University of Iowa (2014) — link
  82. Stock market’s price movement prediction with LSTM neural networks (2017) — link
  83. A deep learning framework for financial time series using stacked autoencoders and long-short term memory (2017) — link
  84. The impact of COVID-19 on emerging stock markets (2020) — link
  85. The day of the week effect in the cryptocurrency market (2019) — link
  86. A Q-learning agent for automated trading in equity stock markets (2021) — link
  87. Technical trading rules in the cryptocurrency market (2020) — link
  88. Mean Absolute Directional Loss as a new loss function for machine learning problems in algorithmic investment strategies — → 2309.10546
  89. Hedging Properties of Algorithmic Investment Strategies Using Long Short-Term Memory and Time Series Models for Equity Indices — → 2309.15640
  90. The M6 forecasting competition: Bridging the gap between forecasting and investment decisions (2024) — link
  91. Machine learning models predicting returns: Why most popular performance metrics are misleading and proposal for an efficient metric (2022) — link
  92. Evaluating multiple classifiers for stock price direction prediction (2015) — link
  93. A performance comparison of machine learning models for stock market prediction with novel investment strategy (2023) — link
  94. An ensemble of LSTM neural networks for high-frequency stock market classification (2019) — link
  95. Predicting Stock Market Indices Using Classification Tools (2019) — link

2 of 103 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.